Designing personalized support experiences in mental health.
A digital mental health case study about what changed when we stopped treating personalization as something set at intake and started designing for the small moments between appointments — where most of the work of staying engaged actually happens.
The work was framed as engagement. It turned out to be relevance.
- RoleProduct strategy · systems design
- DomainDigital mental health
- ScopeEngagement, retention, personalization
- SurfaceProvider workflow + patient experience
Gradual, not all at once. The shape of the problem was a slope, not a cliff.
Intent rarely failed.
Users entered the platform with clear motivation. Drop-off didn't trace back to onboarding or assessment — it accumulated quietly across the weeks that followed.
Personalization wasn't relevance.
Recommendations matched user attributes but rarely matched the moment. People needed support when emotional need surfaced, not when a queue chose them.
Communication compounded.
Every additional notification reduced the next one's signal value. Patients filtered increasingly fast; providers absorbed the resulting follow-up.
rigid pairing · no feedback
behavior-aware comms · self-directed support moments
Four moves followed from this. None are surprising in isolation; the work was in committing to all four at the same time.
Move support into the moment.
Added self-guided exercises, behavior-aware nudges, and short grounding moments available between provider touchpoints.
Replace static cohorts with behavior signals.
Recommendations and communications began responding to observable behavior — context, progression, drop-off shape — instead of to assignment at intake.
Give the user the first move.
Reduced the steps required to request an appointment, message a provider, or surface a user's own assessment results.
Make communication earn its place.
Cut the standing notification cadence; replaced it with messages triggered by behavior changes and unmet needs. Provider follow-up dropped as a result.
The brief was personalization. The work, in practice, was meeting users where they were — and quietly absorbing the cost of doing so.
Self-initiation expands the surface where users can act unsupervised. Each new self-directed move was reviewed against clinical guidance, not assumed safe.
When users act between sessions, providers see less of the work. Surfacing that activity inside the provider workflow became a parallel design problem.
Every reduction in patient friction landed somewhere else.
- Continuity between sessionsusers had something to do in the gap
- Improved engagement & retentionthe platform's primary measured shift
- Faster moment-to-supportfewer steps between need and action
- Reduced provider follow-upself-initiation absorbed the load
- Stronger self-initiationappointment + comms friction down
The pattern that kept appearing wasn't an engagement problem. It was a relevance problem the engagement metrics were measuring sideways.
Once the work shifted from "how do we get users to come back" to "how do we be present when they show up," most of the downstream design followed on its own.
Restraint mattered more than I expected. The most effective moves were the ones that removed something — a sequence, a notification, a step — and let the user's own timing do more of the work.
The same pattern appears in other adaptive systems I work on now — adjacent to therapy, but shaped by many of the same questions about momentum, responsiveness, and trust.